Another piece falls into place on Zuckerberg’s robot puzzle.
On May 1, Bloomberg reported that Meta has completed its acquisition of the robotics AI startup Assured Robot Intelligence (ARI). Eleven hours later, ARI co-founder Xiaolong Wang confirmed the news on X, providing additional first-hand details beyond what Bloomberg reported:
“Excited to share that Assured Robot Intelligence (ARI) has joined Meta to help build the future of humanoid intelligence!”
The entire ARI team will be integrated into Meta Superintelligence Labs (MSL) — Meta’s flagship AI research department, not a peripheral hardware group.
ARI: Founded One Year Ago, Aiming for “Physical AGI”
ARI was co-founded by Xiaolong Wang and Lerrel Pinto roughly a year ago. From the outset, the company positioned itself at the foundation of robotic intelligence — not building complete robots or hardware, but developing the AI foundational models for robots.
Wang’s path is relatively clear: bachelor’s degree from South China Agricultural University → master’s from Sun Yat-sen University → Ph.D. in Robotics from CMU → researcher at NVIDIA (contributing to building foundational large models in robotics) → associate professor at UCSD → ARI co-founder. His research focuses on representation learning from video and physical interaction data — in simple terms, enabling robots to understand the real world by watching videos and learning through trial and error.
Lerrel Pinto is an assistant professor at NYU and a leading scholar in large-scale robotic learning. His research spans data and models, perceptual representation learning, action modeling, and reinforcement learning, with a long-standing commitment to developing low-cost, open-source robotic platforms.
The two have a clear technical roadmap at ARI: train a general-purpose humanoid physical agent, with scaling coming from directly learning human experience rather than relying solely on teleoperation.
This is actually a significant technical judgment. Most current robotics companies rely on teleoperation to collect data — humans wearing VR headsets control robots to perform tasks, generating large volumes of data but with limited generalization ability. ARI’s approach is closer to that of large language models: training a generalizable foundational model using large-scale video and interaction data, essentially bringing the GPT training paradigm into the physical world.
Why Did Meta Buy ARI?
Meta has been laying groundwork in humanoid robots for some time. The Robotics Studio, established in 2025, focuses on the underlying control and hardware technology for humanoid robots. But what Meta lacked was the “brain” — the foundational AI model capabilities for robotics.
ARI fills exactly that gap.
Post-acquisition, ARI’s core mission will be AI model optimization for robotic scenarios, combining their algorithmic capabilities with Meta’s computing power and engineering resources. The Robotics Studio handles hardware and control, while ARI handles perception and decision-making — a complete “body + brain” combination.
Notably, just two days before the acquisition announcement, Meta had significantly raised its 2026 capital expenditure budget to $125–145 billion, primarily directed toward AI data centers and high-end hardware. Zuckerberg has already made it clear that the strategic focus has shifted from the metaverse to AI + embodied intelligence.
A Bigger Signal: The Strategic Pivot of Superintelligence Labs
ARI is joining Meta Superintelligence Labs (MSL), not the Robotics Studio. This detail is worth noting.
MSL is Meta’s flagship AI research department, led by core figures such as Yann LeCun and Joelle Pineau. Placing the robotics team within MSL means Meta isn’t building “a robot project” — it’s constructing the physical embodiment of general superintelligence.
Wang’s own words in his X post: “help bring personal superintelligence into the physical world.”
This differs from the logic behind OpenAI’s acquisition of Figure Robotics (2025, valued at $2.6 billion). OpenAI entered the robotics track by investing in Figure’s whole-machine hardware capabilities, while Meta chose to acquire an AI foundational model team, directly integrating the “intelligence” layer into its own research system.
Two routes: one buys hardware, the other buys a brain.
The Competitive Landscape
Key players in the humanoid robot sector currently include:
- Tesla Optimus: In-house whole machine + in-house AI stack, fastest progress toward mass production
- Figure + OpenAI: Figure builds hardware, OpenAI provides AI — a partnership model
- Boston Dynamics + Hyundai: Traditional robotics company transitioning
- 1X Technologies: Norwegian company focused on home robots
- Unitree: Chinese company, cost-performance approach
- Agility Robotics: Warehouse and logistics scenarios
Meta’s entry adds a competitor on the “AI foundational model” dimension to this race. Meta doesn’t sell robots, but if it builds a general-purpose humanoid intelligence model, it could change the entire sector’s technological trajectory through open-sourcing (Meta’s tradition).
Key Data
- Acquirer: Meta Platforms
- Acquired: Assured Robot Intelligence (ARI)
- Founders: Xiaolong Wang (UCSD), Lerrel Pinto (NYU)
- Lead Investor: AIX Ventures
- ARI Founded: ~1 year ago (mid-2025)
- Integration Target: Meta Superintelligence Labs (MSL)
- Acquisition Price: Undisclosed